All pull requests

Pull requests

• [Fix] mem_cache: Count unified-cache hits only on the finishing insert

  documentationunified-radix-cache
  #44893·lixiufei-leo opened Oct 17, 2026Contributor·5 tasks·

  ·

• [Fix] [H200] ci: Split GLM-5.3-Flash serving recipes into separate tests

  #44892·alphabetc1 opened Oct 17, 2026Collaborator·

  ·
  2

• [Feat] PD: Add opt-in first-chunk decode KV allocation

  documentation
  #44891·sufubao opened Oct 17, 2026·

  ·

• [Perf] [Intel] gemma4: Use exact softcap comparison ranks for greedy verification

  jit-kernelrun-ci
  #44890·alexnails opened Oct 17, 2026Collaborator·5 tasks done·

  ·

• [Feat] PD: Support KV-only hybrid decode radix cache with linear EAGLE

  unified-radix-cache
  #44889·Jzz1943 opened Oct 17, 2026Contributor·

  ·
  1

• [Fix] diffusion: pipeline and model-helper bug fixes

  diffusion
  #44888·OnePunchMonk opened Oct 17, 2026Contributor·

  ·

• [Fix] diffusion: runtime utils bug fixes

  diffusionMulti-modal
  #44887·OnePunchMonk opened Oct 17, 2026Contributor·

  ·

• [Fix] kda: Align graph tracking branch dtype with the recurrent state pool

  jit-kernel
  #44886·Jzz1943 opened Oct 17, 2026Contributor·

  ·

• [Perf] [Intel] gemma4: Fuse greedy verification logits finalization

  jit-kernelrun-ci
  #44884·alexnails opened Oct 17, 2026Collaborator·5 tasks done·

  ·

• [Feat][NPU] enable device-side multimodal feature transport via torch_npu IPC

  npu
  #44883·sanqiannianyiqian opened Oct 17, 2026·5 tasks done·

  ·
  1

• Fix stale sgl-kernel paths in Metal kernels README link labels

  documentationsgl-kernel
  #44882·aarushkandukoori opened Oct 17, 2026·

  ·
  1

• [CI] base-b: Move kernel and unit tests that pass on the 5090 to 1-gpu-small

  blackwellhicachememory-poolquant
  #44880·mmangkad opened Oct 17, 2026Collaborator·

  ·
  4

• [Perf] deepseek_v4: write the indexer's page slots with one kernel per prefill layer

  jit-kernel
  #44879·kpham-sgl opened Oct 17, 2026Collaborator·

  ·
  2

• [Perf] moe: apply the routed scaling factor inside the fused gate for FP8 MegaMoE

  #44878·kpham-sgl opened Oct 17, 2026Collaborator·

  ·
  2

• [CI] base-b: Move server tests that pass on the 5090 to 1-gpu-small

  hicacheloraspeculative-decoding
  #44876·mmangkad opened Oct 17, 2026Collaborator·

  ·
  6

• [Perf] deepseek_v4: build the DeepGEMM prefill indexer geometry once per forward

  deepseek
  #44875·kpham-sgl opened Oct 17, 2026Collaborator·

  ·
  2

• [Fix] scheduler: Cap aborted speculative output at max_new_tokens

  speculative-decoding
  #44874·YueshenZ opened Oct 17, 2026·4 tasks done·

  ·
  1

• [Feat] mem_cache: defer prefill for pending external linker lookups

  hicacheunified-radix-cache
  #44873·feichai0017 opened Oct 17, 2026·3 of 4 tasks·

  ·

• [Fix] request_logger: truncate list elements in --log-requests output

  #44871·LibertychaserUS opened Oct 17, 2026·2 tasks done·

  ·
  1

  1

• [Perf] [Intel] gemma4: Fuse QKV normalization and RoPE

  jit-kernelrun-ci
  #44869·alexnails opened Oct 17, 2026Collaborator·5 tasks done·

  ·

• [Fix] deterministic: Disable the hybrid SWA memory pool with the radix cache

  #44868·aarnetalman opened Oct 17, 2026·4 of 5 tasks·

  ·

• [Fix] scheduler: Count retracted output tokens as uncached HRRN work

  #44867·luicarus opened Oct 17, 2026·3 of 5 tasks·

  ·

• [Feat] speculative: Online acceptance estimator for DSpark confidence-scheduled verification

  deepseekdocumentationjit-kernelspeculative-decoding
  #44866·maocheng23 opened Oct 17, 2026Collaborator·3 tasks done·

  ·

• [Diffusion] ComfyUI: online INT8 quantization for BF16 Qwen-Image-2.1 single files

  diffusiondocumentation
  #44865·OnePunchMonk opened Oct 17, 2026Contributor·

  ·

• [Perf] DSv4.1: Enable encoder bounded replay with shared BCG metadata

  deepseekmemory-pool
  #44864·Oasis-Git opened Oct 17, 2026Collaborator·3 of 4 tasks·

  ·
  3